Abundant intelligence, scarce judgment.

Eleven leaders, rebased to 100

From 30 September 2021 to 30 September 2026, Nvidia's share price rose about elevenfold and the median of the eleven companies rose to 188. Salesforce was the only one of the eleven below its starting level, and Tencent was about flat in dollar terms. Almost every company hit its low in late 2022, when rising rates cut technology valuations. Oracle, SAP and IBM are the exceptions: their deepest falls came in 2026, in the selloff of software and heavily indebted AI-infrastructure names.

Within each peer group the spread is wider than between groups. Among hyperscalers, Alphabet (255) pulled ahead of Microsoft (188) and Amazon (150), while Oracle rose to about 340 in September 2025 on AI-cloud contracts and then gave back half of it. In enterprise software and services, IBM (212) beat SAP (170) and Salesforce (85). Among consumer platforms, Apple and Meta ended close together, but Meta got there from a 73% drawdown in 2022; Tencent ended where it started.

The drawdowns change the picture more than the endpoints do. Meta fell 73% before recovering to 221. Oracle ended at half its September 2025 peak. IBM, SAP and Oracle had their deepest drawdowns in 2026, not 2022.

Drawdowns use month-end closes, so falls within a month were deeper than shown. Recent returns computed from the index were cross-checked against companiesmarketcap.com for five companies; the gaps of one to five points come from dividend adjustment and the time of day each source uses.

What an agentic organization really is

An AI agent is an application that pursues a goal by observing its environment and acting on it through the tools at its disposal: systems, data, other agents. The model supplies the reasoning. Everything around it, the goal, memory, tools, guardrails and the point where a human is called in, determines whether it is reliable. An agentic organization is one in which such agents become regular participants in workflows rather than tools occasionally consulted.

Agents change how work is coordinated, because tasks can now move across the company without passing through a reporting line. In 2023, Maurício Pontuschka and I argued in "As organizações do futuro" that the evolution of computing systems has always moved in step with the evolution of organizations, and that both can be read through the same three lenses: structure (the elements and how they are arranged), behavior (what each element does to produce value) and dynamics (how the elements interact to serve the whole). Organizations moved from vertical structures centralized in corporate functions, to multidivisional structures organized by region and segment, to matrix structures that share responsibility between functions and business units, and now to a range of experiments with decentralization: adhocracy, holacracy, decentralized autonomous organizations and what we called organizational biomes. Computing moved from monolithic programs through networks, graphical interfaces, the internet, mobile and cloud to artificial intelligence, with quantum computing ahead.

The shared arc is the point. Both systems began with structural and behavioral elements mixed in a single model, then separated them into layers, and finally turned to modeling dynamics: how processes interact with one another and with the outside world. When computing moved to modeling dynamics, organizations had to follow. Agents are that step for organizations, because they act directly on all three lenses at once. They redraw who does what, move people up from execution to specification and review, and coordinate work at machine speed.

The ambition is not new. In BMI's REBOOT study , 92% of the organizations analyzed ran matrix or multidivisional structures, while 58.7% aspired to network models. Agents may finally supply the connective tissue that network organizations always lacked. But connection is not the same as coordination, and coordination is not the same as commitment. Most organizations can redesign structure and specify behavior; dynamics, where the tacit norms, influence and pacts between teams live, is where most AI programs stall. We also wrote that there is always a time lag between a technological advance and its effective absorption by organizations. As Christopher Smart observed of AI adoption, "The horses need to drink it."

It's not just about the best water or the nearest water. The horses need to drink it. Christopher Smart , "The Most Important AI Race", August 2026

High tech, high touch: the case for Essential Leadership

I have long summarized my view in a simple formula: high tech, high touch. The higher the technological standard of an organization, the higher its humanistic standard of leadership must be. This is not a trade-off between efficiency and humanity. It is the condition for efficiency to last. When an agent executes thousands of decisions a day, the values, criteria and blind spots of the people who designed it are multiplied thousands of times. Technology does not dilute leadership's responsibility. It amplifies it.

This is the core of what I have called Essential Leadership, in my book A Liderança Essencial and in BMI's work on the theme: timeless principles such as work ethic, transparency and strategic vision, combined with the capabilities a networked world demands, namely collaboration across networks, learning agility, emotional intelligence and holistic health. In an agentic organization, the timeless principles become the specification for the machines. The new capabilities become the conditions for people to thrive alongside them.

High touch also has a social and ethical dimension, because the agentic transition is at the same time a social transition. The International Monetary Fund estimates that about 40% of global employment is exposed to AI and warns that inequality may rise if the gains accrue mainly to high-income workers and to capital. For a corporation, this becomes concrete: telling employees honestly which roles will change, investing in reskilling before displacement rather than after, and refusing to delegate decisions that touch people's dignity, such as hiring, dismissal and performance evaluation, without meaningful human review. People already sense the stakes. In Afferolab's research on AI as an agent, half of the executives surveyed said they were uncomfortable with AI evaluating performance.

The higher the technology, the higher the touch. Daniel Augusto Motta

Where the human touch creates value

The human touch is not a soft variable. It shows up in three places where value is created or destroyed: trust, relationships and meaning.

Trust is where value is unlocked. Any task whose output must be checked by a person retains most of its original cost; value appears at the threshold where verification can stop. That threshold is crossed not by a model release but by a human decision to rely on the system, backed by evidence. Customers extend trust to companies, not to algorithms, and a single agent error in front of a client can undo years of reputation.

Trust is also designed, and sometimes that means adding effort rather than removing it. Renée Richardson Gosline has argued in Harvard Business Review that AI customer journeys need more friction , not less: bad friction, such as confusing forms or journeys that are easy to enter and hard to leave, should go, while beneficial friction, a deliberate pause before a consequential or irreversible choice, keeps human judgment in the decision. Her experiments point the same way inside the firm. Highlighting likely errors in AI-generated drafts improved accuracy without significantly lengthening the task , yet participants did not believe the prompts had helped: they overestimated their own ability to spot the machine's mistakes. Her conclusion is a design principle for any agentic organization: beware of individual-level solutions for structural problems. High touch is not asking every person to be vigilant. It is building the speed bumps into the system, at the moments that matter, and making exit as easy as entry.

The pressure runs in both directions. Leaders are pushed to deploy agents in service, yet in a Gartner survey 64% of customers said they would prefer companies not to use AI for customer service . The answer is not to hide the machine, but to decide deliberately where it acts and where a person steps in, and never to lose sight of what the customer is actually buying. Kodak built one of the first digital cameras in 1975 and filed for Chapter 11 in January 2012 ; three months later Facebook agreed to pay US$1 billion for Instagram , a company of 13 people. Customers never wanted film. They wanted pictures, shared. The same is true of AI: the tool matters only as far as it serves the relationship.

Relationships are where AI amplifies people rather than replacing them. In the largest field study of generative AI at work, Brynjolfsson, Li and Raymond found that an AI assistant raised the productivity of 5,179 customer-support agents by 14% on average and by 34% for novice and less-skilled workers. The tool worked by spreading the tacit practices of the best agents to everyone else. The machine carried human know-how; it did not substitute for the human conversation.

Meaning is where the risk hides. At the House of Brains research center we have studied what happens when performance pressure outruns human limits. In our Felicidade Artificial study, an exploratory survey of 62 respondents, 74.2% agreed that a manager's own behavior pressures subordinates to follow, and 56.5% named the pursuit of high performance as a trigger for unhealthy coping. In The Burnout Economy , which I presented at SXSW 2026, I described leaders as biological amplifiers and warned that the next burnout wave will not come from working harder. It will come from never finishing. Agents never sleep, and an organization that lets their pace set the human pace will burn out the very people it needs to supervise them.

Designing the human-agent operating model

High touch begins with design. An operating model that adds agents is a decision about where people stand in the flow of work, and most companies make it by default, one tool at a time. The result is predictable: people end up where the machine left gaps, not where their judgment, empathy and relationships add most.

A useful division of labor follows from what each side does best. AI is extremely fast and holds a large short-term memory, but it lacks the full context of the business and struggles with creativity across many steps. People bring a lifetime of context and the ability to connect ideas no dataset links. Put AI on well-defined tasks, planning and execution; put people on objectives, guardrails, exceptions and strategy. Asking people to rubber-stamp routine approvals wastes the human advantage and burns out the people who hold it.

The practical rule is simple. Every agent needs a named human owner. Every consequential action needs a defined point where a person approves, and a record of what the agent did and why. Autonomy is extended one step at a time, as evidence accumulates, never granted in a single leap because a vendor demo was impressive.

Structures will move under this pressure, and that is precisely where the human touch is most at risk. Gartner predicted in 2024 that through 2026, 20% of organizations would use AI to flatten their structure , eliminating more than half of their current middle-management positions. Middle managers are not only coordinators that agents can replace. They are the carriers of culture, the translators of strategy and the first line of care for people. An organization that removes them must redesign deliberately who will perform those human functions, or it will discover their value only after it has lost them.

Culture, learning and the people agenda

Agents inherit their makers. As Afferolab put it in its research on AI as an agent , every AI agent carries human traits in its code: our criteria, our haste, our gaps and our biases. An organization that is careless in how it decides will build careless agents, and then run them at machine speed. Alessandra Lotufo, who leads Afferolab, summarized it in one line: AI does not fix chaos; it amplifies it.

Culture is also where the digital organization differs most from its predecessors. In Anthesis I argued that digital organizations share three cultural drivers that reinforce one another: they are light, empathetic and organic. Light means abandoning the premise that control and the accumulation of capital are the long-term levers of value; operating models are rebuilt to create value through access rather than ownership, and committed relationships give way to situational partnerships in dynamic networks. Empathetic means that the only valid perspective is the experience of the other: customer, user and employee experience combine into a single view of value centered on people. Organic means abandoning the logical, planned and linear premise of the industrial paradigm; when resources, capital included, become abundant, value moves to orchestrating them into collaborative, ecosystem-based delivery.

The drivers form a cycle. Lightness frees an organization from what it owns, so it can look outward at the experience of others. Empathy reveals needs that no single firm can meet alone, which pushes the organization toward organic delivery. And orchestrating an ecosystem is possible only for an organization light enough to connect and disconnect quickly. AI tests each driver. It makes organizations lighter, but only if the context, data and decision records that agents depend on are built deliberately: lightness of structure demands depth of knowledge. It makes empathy scalable, and organizations that forget the employee side, how their own people experience working alongside agents, will find that use goes underground. And it completes the organic premise. When intelligence joins capital among the abundant resources, value sits more than ever in orchestration, and orchestration is a human act.

Three cultural conditions matter most. The first is psychological safety, in Amy Edmondson's classic sense : people must feel safe to say "the agent is wrong" or "I don't understand why it did that" without being seen as resisting progress. Oversight dies in silence. The second is learning that changes behavior. In BMI's Design Mastery study of more than 100 CHROs, 62% prioritized digital upskilling, yet 56% said their upskilling programs showed no measurable progress. Training people to prompt is not the same as preparing them to supervise, specify and decide. The third is protected recovery. If agents run continuously, humans must not.

A fourth condition is less visible and may matter most over a decade: the transmission of tacit knowledge. Michael Polanyi's observation that "we can know more than we can tell" describes the part of an organization no model can absorb: how a seasoned negotiator reads a room, how a senior partner senses a client's real concern, how an engineer recognizes a design that will fail. That knowledge has always passed through apprenticeship, in the drafting, research and first-pass analysis that juniors did beside experienced colleagues. Agents now do that work, and companies are drawing the obvious conclusion. Using payroll data from the largest US payroll provider, Erik Brynjolfsson, Bharat Chandar and Ruyu Chen found a 13% relative decline in employment for workers aged 22 to 25 in the most AI-exposed occupations , while more experienced workers in the same occupations held steady or grew.

The saving is immediate; the cost is deferred. An organization that stops forming juniors looks highly productive today, carried by a shrinking group of experienced people using AI, and may discover years later that nobody left knows how to do the work. Junior hiring belongs with research and development, as an investment in a capability pipeline rather than a cost to optimize away. The apprenticeship itself has to be redesigned: juniors attempt the analysis first and then compare it with the agent's, critique AI output instead of accepting it, sit in the client meetings and negotiations where judgment becomes visible, and receive the frequent feedback no agent can give. Senior people must be measured and rewarded for forming others, or they will not do it. What can be captured should be captured, through decision records, after-action reviews, experts explaining their reasoning and structured interviews before retirement, and fed into the context agents use. It will never be all of it. That is precisely why the human channel must stay open.

The same study identified three roles for the modern CHRO: architect, choreographer and articulator. They fit the agentic organization well. The architect designs how humans and agents divide the work. The choreographer sets the rhythm so that speed does not become exhaustion. The articulator keeps the relationships, inside and outside the firm, that no agent can hold.

What boards and CEOs should ask

Andrej Karpathy has observed that large language models benefit individuals disproportionately , while their impact on corporations and governments lags. I read that less as a verdict on technology than as a verdict on organizations. Individuals adopt what helps them. Institutions adopt only what their structures, incentives and cultures allow. The agentic organization will be won or lost in that gap.

For boards and CEOs, the human touch translates into five questions worth asking at every AI review. Which decisions are agents allowed to take, and who approved that boundary? Who is the named owner of each agent in production? Where has saved time actually gone? Can our people tell us, without fear, when an agent is wrong? And are our leaders setting a pace that humans can sustain while machines run through the night?

High tech, high touch is not nostalgia. It is a design principle. The corporations that thrive in the agentic era will be those that let machines do what machines do best, precisely so that their people can do more of what only people can: earn trust, hold relationships and give work its meaning. The more advanced the technology, the more essential the leader.

What the evidence actually shows

The evidence is strong at the level of the task, modest at the level of the economy, and uneven at the level of the firm.

At the task level, controlled studies are convincing. Customer-support agents with an AI assistant resolved 14% more issues per hour , and 34% more among novices. Consultants working inside the model's capability frontier produced more than 40% higher-quality work , though they did worse on tasks just outside it. At the level of the labor market, the picture is more sober. The Federal Reserve Bank of St. Louis estimates, from self-reported data, that generative AI saved 2.2% of all work hours by mid-2026 , up from 1.6% in 2024. A large Danish study by Anders Humlum and Emilie Vestergaard, linking surveys to administrative records, finds precise null effects on earnings and hours , ruling out effects larger than 2% two years after ChatGPT's launch, even as tasks are being reorganized. Where firm-level returns do appear, they tend to come from growth rather than cost: Babina, Fedyk, He and Hodson found in the Journal of Financial Economics that AI-investing firms grew sales, employment and market value faster , primarily through product innovation, with the gains concentrated in larger firms. The broadest firm-level evidence is sobering: in a survey of nearly 6,000 senior executives across the United States, the United Kingdom, Germany and Australia, run with the Atlanta Fed, the Bank of England and the Bundesbank, about nine in ten reported no impact of AI on productivity or employment over the previous three years, while expecting a gain of about 1.4% over the next three.

Some widely quoted numbers deserve caution. The claim that 95% of organizations see no return on generative AI came from a report based on 52 interviews and 153 survey responses, and has drawn serious methodological criticism . Pessimistic statistics can be as unreliable as vendor promises.

This is why I encourage boards to scale capital with the level of evidence reached, not with the level of excitement in the room.

The industry's ROIC question

Return on invested capital is the discipline that separates a technological revolution from a good investment. It asks whether the after-tax operating profit generated by the capital deployed exceeds what that capital costs. For the AI buildout, the invested capital is now visible. For 2026, Amazon has guided to about US$220 billion of capital expenditure, Alphabet to US$195–205 billion , Meta to US$130–145 billion and Microsoft to roughly US$175 billion for the calendar year.

Measured more broadly, the bill is larger still. Goldman Sachs forecasts global AI investment above US$1 trillion in 2026 , US$581 billion of it in the United States, about 1.8% of US GDP and rising, and argues that the widely quoted hyperscaler consensus understates the total because it leaves out other public companies and private players.

Today, the clearest returns sit with the suppliers. Nvidia reported data-center revenue of US$89.0 billion in a single quarter, at a 75% gross margin . The buyers' revenues are growing fast but from a lower base. Microsoft reported that its AI business surpassed a US$37 billion annual run-rate, up 123% ; Google Cloud grew 82% in the second quarter of 2026, with a US$514 billion backlog; Anthropic's run-rate reportedly reached US$65 billion by the end of July 2026 . These are remarkable numbers. They are still small next to what the capital requires: MIT Technology Review's David Rotman puts the largest hyperscalers' spending at around US$750 billion a year against AI revenues of US$150–200 billion . Gary Gensler, former chair of the US Securities and Exchange Commission, calls the build-out "a parlay bet by the capital markets": it pays only if every leg wins. In the same article, Wharton's Jessica Wachter estimates that AI companies must raise their own productivity 2.7 times to break even by 2030. In 2024, David Cahn of Sequoia framed the gap as "AI's $600B question" ; in July 2026 he updated it to a US$1.5 trillion question , estimating that each year of data-center investment requires roughly twice its value in lifetime revenue.

The challenge is that the spending does not have commensurate revenues yet. Gary Gensler , former SEC chair, via MIT Technology Review , September 2026

The macroeconomic footprint is already visible. The Federal Reserve Bank of St. Louis estimates that AI-related investment contributed 0.97 percentage points of US real GDP growth in the first nine months of 2025 , about 39% of the total and more than the dot-com peak, although the share falls sharply once imported chips and servers are netted out. The financing is changing character as well: global venture funding reached about US$510 billion in the first half of 2026, roughly four-fifths of it to AI , and Big Tech has roughly doubled its debt to fund the build-out.

A simple illustration, using my own arithmetic: if the four largest spenders invest around US$700 billion a year for three years, AI-related invested capital approaches US$2 trillion. Earning 10% after tax on it, roughly the cost of capital plus a modest spread, would require about US$200 billion a year of incremental operating profit, after depreciating hardware that may need replacement within five or six years.

History offers a sobering scale. Stijn Van Nieuwerburgh estimates for the Brookings Institution that the AI pipeline will trigger about US$10.3 trillion of investment from 2025 to 2032 , an average of 3.63% of US GDP a year, well above the railroads at 2.24% and more than three times the telecom and fiber boom. In every earlier boom, the economy gained from the infrastructure; investors did not always. Part of the bill is also less visible than capex. The Wall Street Journal reported that nine large technology companies carry roughly US$3 trillion of commitments off their balance sheets , in leases not yet started and purchase agreements for chips, networking and power, about five times their capital spending of the past year and roughly three times the US$248 billion of lease liabilities and US$356 billion of debt they carry on their balance sheets . And the model layer is burning cash at a scale without precedent: according to projections reported by Bloomberg from the Financial Times , OpenAI expects about US$278 billion of negative free cash flow from 2026 to 2030, even as revenue grows from US$36 billion to US$350 billion a year. As Reuters summarized the same projection, cumulative revenue of about US$840 billion over the period would still fall short of roughly US$856 billion of spending on computing and infrastructure alone.

The pace matters as much as the size. Torsten Slok, chief economist at Apollo, estimates that data-center capex will reach about 3% of US GDP in 2027–29 , two and a half times the telecom peak of 2000, and that it is building at roughly twice the pace of the 2002–05 housing boom. A cycle that builds that fast can unwind just as fast. The debate among financial institutions has shifted accordingly, from demand to financing. J.P. Morgan Asset Management estimates that a US$5 trillion build-out needs about US$650 billion of annual revenue to earn a 10% return , roughly US$35 a month from every iPhone user. Morgan Stanley sees a US$1.5 trillion external financing gap through 2028, much of it to be filled by private credit. Man Group notes that lenders assume asset lives of 10 to 20 years while accelerators become obsolete in a fraction of that time . Revenue, meanwhile, has surprised on the upside. The risk is no longer that it never arrives; it is that debt, short asset lives and concentrated counterparties turn a disappointment into a credit event.

Two warning signs deserve board attention. The first is depreciation. Hyperscalers depreciate servers over five to six years, and Amazon shortened part of its fleet from six to five years in 2025 , citing the pace of AI. If the economic life of accelerators proves shorter than their accounting life, a question still open, reported returns overstate real ones. The second is cash. Capital intensity has outrun operating cash flow: Meta generated only US$784 million of free cash flow on US$60.8 billion of quarterly revenue, and Microsoft is now the only one of the four with positive trailing free cash flow . Debt financed about 32% of big-tech capex by mid-2026, against 9% two years earlier .

My reading is that industry ROIC will compress before it recovers, and that the recovery will be deeply uneven. Returns accrue first to the layers with pricing power. Whether scale gives the cloud and model layers enough pricing power to justify this capital is the central strategic question of the industry.

A framework I have used with boards for years explains why. The centre of gravity of economic value moves through structural layers. Value first concentrates in natural resources. Resources are transformed into products, which capture the larger share. A layer of services is then added on top of the product offer. Platforms follow, built to orchestrate and integrate products and services. Finally, some technologies become true operating systems of human activity, layers on which whole economies run. In Anthesis I put the sequence in historical terms: services added value to products in the mid-twentieth century, platforms integrated services and products at its end, and we would increasingly live with operating systems that consolidate platforms into a complex ecosystem, with AI and cloud computing as the likeliest candidates.

The mechanism is what matters for ROIC. When a new layer emerges, it becomes the new centre of gravity of economic value, and in doing so it exhausts the value previously concentrated in the layer below. Over time, returns on capital in the lower layers converge to their cost of capital, while asymmetric returns concentrate in the upper layers. Nothing disappears: the lower layers remain essential. They simply stop earning more than they cost. For an economist this is the dynamic of economic rents made visible. Rents in a lower layer attract entry and imitation until margins normalize; the upper layer earns rents because it controls the scarce resource of the moment, whether access to customers, the power to integrate or the rules of the system itself.

AI is running the sequence at unusual speed and scale. Energy, chips and data centers are the resource layer, and they absorb most of the capital described above. Models behave like products and are already commoditizing. Services built on models are multiplying, agent platforms and open protocols are forming to orchestrate them, and the contest is over which of them becomes the operating system of knowledge work. The framework points to the hard question for investors. If capital is concentrated in the lower layers while asymmetric returns migrate upward, part of today's investment will earn no more than its cost of capital, however essential the infrastructure turns out to be.

The battle, layer by layer

The AI industry is a stack of five layers, and each is being fought over with different weapons. The prize is enormous: Gartner forecasts worldwide AI spending of US$2.67 trillion in 2026 , of which about US$1.48 trillion goes to AI infrastructure, and its analysts describe the buildout of AI data-center capacity as the largest infrastructure project humanity has undertaken.

Lithography and foundry. At the bottom, the market is already a monopoly in all but name. ASML is the only company that makes the extreme-ultraviolet lithography machines needed to manufacture every advanced processor. One level up, TSMC held 72.5% of global foundry revenue in the second quarter of 2026 , with its most advanced capacity fully booked by demand for AI chips. Two companies, one in the Netherlands and one in Taiwan, are single points of failure for the whole industry. Beneath both sits energy: the International Energy Agency expects data-center electricity use to roughly double to about 945 TWh by 2030 , and BloombergNEF projects that US data centers will take a fifth of the country's power by 2035 . Access to power is becoming a competitive weapon in its own right.

Accelerators. Nvidia reported US$89.0 billion of data-center revenue in a single quarter, at a 75% gross margin . Its moat is software as much as silicon: CUDA, which Jensen Huang said in 2024 had 5 million developers . The challengers are Nvidia's own largest customers. Google's TPUs now anchor multi-gigawatt commitments with Anthropic and Broadcom from 2027 , and Amazon's Project Rainier runs on nearly half a million Trainium2 chips . The contest is between a general-purpose standard and vertically integrated custom silicon, with geopolitics on top: in January 2026 the US moved exports of advanced accelerators to China to case-by-case licensing .

Custom silicon moves the toll gate rather than removing it. Silicon Analysts, which maps 52 companies across ten layers of the AI hardware chain, finds pricing power concentrated at three chokepoints : Nvidia's accelerators, anchored by CUDA; custom-chip co-design, where Broadcom and Marvell hold roughly 95% of the market; and interconnect and optics, because AI racks need ten to thirty-six times more fiber than conventional ones. Custom chips from the hyperscalers cost 40–65% less to run at scale, yet carry only 15–20% of those companies' own workloads. Almost every major player is at once customer, supplier and competitor of the others.

Cloud. Three firms dominate. Synergy Research estimates that in the second quarter of 2026 Amazon held 28% of cloud infrastructure services, Microsoft 20% and Google 15% , in a market that grew 43% to US$143.4 billion in the quarter. The same firms finance and distribute the model labs, forming what the UK Competition and Markets Authority called an "interconnected web" of more than 90 partnerships and investments .

Models. The price of staying at the frontier keeps rising. Epoch AI estimates that the cost of frontier training runs has grown about 2.4 times a year since 2016 , with the largest runs expected to exceed US$1 billion by 2027. Independent benchmark organizations now referee the contest. In September 2026, the top positions on both the LMArena text leaderboard and the Artificial Analysis Intelligence Index were held by models from Anthropic, OpenAI and Meta, separated by margins close to statistical noise.

Distribution. Finally, the battle for the user. ChatGPT passed 900 million weekly users in early 2026, and Alphabet reported 950 million monthly users of the Gemini app in July. Whoever owns the daily interface owns the knowledge of what people ask, and the default position from which every other service is reached.

Why the market may tip toward one or two winners

Economists call an industry a natural monopoly when a single firm can serve the whole market at lower cost than several competitors could, typically because fixed costs are enormous and marginal costs are low. The frontier of AI increasingly shows these features. Anton Korinek and Jai Vipra conclude in Economic Policy that AI displays "significant economies of scale and scope that may create a tendency towards greater market concentration." Five mechanisms reinforce one another.

Significant economies of scale and scope that may create a tendency towards greater market concentration in the future. Anton Korinek and Jai Vipra , "Concentrating Intelligence", Economic Policy , 2025

1. 1.Fixed costs. Billion-dollar training runs and multi-gigawatt data centers can be financed by very few firms on earth.
2. 2.Low marginal costs. Once a model is trained, serving an additional customer costs a small fraction of the investment, so the largest player can always undercut.
3. 3.Feedback loops. More users generate more interactions and more feedback, which improve the next model and attract more users.
4. 4.Switching costs. Developer tools, agent integrations and enterprise contracts lock customers in. The US Federal Trade Commission found that AI partnerships between cloud providers and labs "may increase contractual and technical switching costs" .
5. 5.Vertical integration. The leaders are integrating from chips to cloud to models to consumer applications, capturing margin at every layer.

Vertical integration now runs through balance sheets as well as products. Bloomberg's guide to the circular deals underpinning the AI boom shows that the largest investors in the frontier labs are also their largest suppliers: Nvidia, Microsoft, Amazon and Google each hold equity in OpenAI or Anthropic while selling them compute. Part of a supplier's reported demand is therefore financed with its own capital, and a reset in private valuations would shrink commitments across the whole web at once. The board test for any AI supplier or partner follows directly: what share of its backlog comes from counterparties it has itself financed?

There are real counterforces. The Stanford AI Index 2025 showed the gap between the first and tenth models on Chatbot Arena narrowing from 11.9% to 5.4%, and the gap between the best US and Chinese models from 9.3% to 1.7%. DeepSeek's R1 wiped close to US$600 billion off Nvidia's market value in a day in January 2025, and open-weight models handled roughly a third of tokens routed through OpenRouter by late 2025. The Stanford AI Index 2026 shows both forces at once: industry produced more than 90% of notable frontier models in 2025, yet the six leading labs sat within 25 Elo points of one another, and the top US model led the top Chinese model by only 2.7%. The plumbing is opening up too. Anthropic donated the Model Context Protocol, the standard that connects agents to tools and data, to a neutral foundation ; by December 2025 it counted about 97 million monthly SDK downloads. Open plumbing with closed models moves value toward distribution, infrastructure and proprietary data. Antitrust authorities are watching closely.

Epoch AI's tracking shows the shape of that convergence. On its composite Capabilities Index, the best open-weight models have trailed the best closed ones by about eight points, roughly four months of frontier progress , since the start of 2026. Open models catch the last generation, not the current one, and the open frontier is now largely Chinese. Frontier pricing power rests on those four months, and it holds only for work that needs them.

My reading is that these counterforces commoditize the middle of the market, not its frontier. Capable models will become cheap and abundant. The frontier itself, meaning the most capable models, the largest pools of compute and the default consumer interface, tends to converge toward one or two global winners, possibly one per geopolitical bloc. For corporations, the planning assumption should be concentration of supply, not permanent competition among many equals. That is exactly why the hooks matter.

When you go through a curve like this, the risk of underinvesting is dramatically greater than the risk of overinvesting for us here. Sundar Pichai , CEO of Alphabet, Q2 2024 earnings call

Concentration of supply also changes the rules for everyone who competes on top of it. For some years I have refined the idea of a Digital Competitive Arena to describe these rules, organized around two principles: always on, and always beta. The traditional pursuits of market share, control of distribution channels, long research cycles, legal protection of intellectual capital and heavy investment in physical assets are all put in check. As I wrote in Anthesis , share of wallet becomes more relevant than product segmentation, hub centricity overtakes vertical integration, and user experience catalyzes value along fluid journeys through the value chain. Hub centricity is the position of the node through which interactions must pass: the platform, marketplace or interface that connects many parties and learns from all of them. User experience is the quality of the journey, with friction removed and needs anticipated. Share of wallet and attention is what is at stake: the proportion of customers' budget and, increasingly, of their time that a company captures, against competitors from any category. At the centre sits the condition that holds the arena together. Digital competitors never finish a product; they release, measure and adjust continuously, and an offer that stops iterating starts to lose position on all three vertices at once.

AI changes each vertex. The hub is moving to the agent. When an assistant chooses the supplier, books the trip or drafts the contract, the agent becomes the node every transaction passes through, and brands compete for the agent's recommendation rather than the customer's glance. Meta has just paired its Muse agent with a small wearable companion device , part of a bet that agents, not feeds or app stores, become the front door to digital life and commerce. User experience shifts from screens to conversations and delegated outcomes. Attention stops being purely human, because a growing part of every purchase journey is read, compared and filtered by machines. And always beta now applies to strategy itself, because the underlying capability changes every few months. For a board, the question is no longer "what is our digital channel?" It is "whose agent will recommend us, and why?"

Four hooks to watch

The first is the price hook. Unit prices keep falling: the inference cost of GPT-3.5-level capability dropped more than 280-fold between late 2022 and late 2024. But consumption rises faster, because agents and reasoning models use many times more tokens per task. The FinOps Foundation captured it: "A token may get cheaper; tokens, in aggregate, are not." As the early subsidy phase ends, procurement shifts from counting seats to committing to future consumption. A company that has not instrumented its own AI costs will negotiate blind.

The second is the context hook. Models are rented; context is owned, but only if you keep it. Every prompt, document, decision and rejected alternative an agent processes is the raw material of your future advantage. If that record lives only inside a vendor's platform, in formats you cannot export, you are renting back your own memory.

The third is the workflow hook. AI features embedded in software subscriptions aggregate token consumption behind a seat price. Visibility disappears, and costs drift upward at every renewal. Once agents are woven into daily workflows, switching becomes a reorganization, not a procurement decision.

The fourth hook is the most subtle, because it is set inside our own heads. A Microsoft Research and Carnegie Mellon study found that higher confidence in generative AI was associated with less critical thinking among knowledge workers. An MIT Media Lab preprint , small and not yet peer-reviewed, reported weaker neural engagement among writers who relied on a language model. The evidence is early. The risk is plausible enough to manage: an organization that outsources its thinking eventually loses the capacity to judge whether the thinking it buys is any good.

The hook is not a matter of individual weakness, which is why exhortation will not remove it. Eric So, in his forthcoming book The Collision: What AI Does to Us , describes a gravity built from four forces: a brain that consumes about a fifth of the body's energy and is wired to conserve effort; contests in which rewards are relative and concentrated at the top, so that even a small edge decides; a prisoner's dilemma, because AI-assisted work is hard to detect and abstaining looks like falling behind; and the plain fact that the tools do raise productivity. Telling people to think for themselves ignores all four. Once use begins, a second mechanism takes over, a downward spiral in which more AI use means less practice, skills weaken through disuse, and the growing discomfort of working without AI drives still more use. Nor is the direction of human capability guaranteed. Bernt Bratsberg and Ole Rogeberg showed in PNAS that the long rise in measured IQ reversed for Norwegian cohorts born from the mid-1970s, and because the decline appears among brothers in the same families, its causes are environmental. AI did not cause that reversal. It does show that cognitive capability responds to the environment, and a technology that offers to do the thinking is a large change to that environment. Because the pressure is systemic, the response must be too. An organization that rewards only speed and volume has already chosen the spiral.

Why these five skills, and why now

Agents are moving from pilots into the core of work. Gartner predicts that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028 , up from none in 2024, and that a third of enterprise software applications will include agentic AI by then. In the same analysis, Gartner warns that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls, and estimates that only about 130 of the thousands of vendors claiming agentic capabilities are real. The distance between that promise and those cancellations is not a technology gap. It is a leadership gap.

The gap has a deeper cause. Two transitions are rewriting the premises of leadership at the same time. The first is social. In A Liderança Essencial I described the passage from an ethic of duty to an ethic of pleasure as a revolution in the social norm: a shift toward satisfying individual needs, expectations and aspirations, at the expense of obedience to family, religious, political and institutional structures of power. Work and consumption change structurally in the search for experience, expression and impact, and the psychological contract between individual and organization has to be rewritten. The duty ethic organized work around norms and standards, command and control, and centralization, and it bound people to organizations through loyalty, stability and hierarchy. The pleasure ethic organizes work around fluidity, autonomy and interdependence, and the bond now comes from connection, expression and experience.

The second transition is technological. Digital transformation is not one more application of technology; its power lies in convergence, led by artificial intelligence together with the industrial internet, robotics, extended reality and genomics, and sustained by data analytics, cloud computing and social networks. For organizations built on the old paradigm, as I wrote in 2019, it is "a question of renewed life or announced death". In work relations, the pillars move from predictability, incrementalism and departments to volatility, rupture and ecosystems. The premises of effective design move from cohesion, perfection and self-sufficiency to connectivity, collaboration and permeability.

The two transitions reinforce each other, and AI accelerates both. Agents are permeability in its purest form: capability that enters the organization from outside, on demand, across every boundary. And professionals choosing their own AI tools, on their own terms and outside the official channel, are the pleasure ethic at work. Leaders who answer both with the instruments of the old paradigm, prohibition, centralization and appeals to loyalty, push the most valuable experimentation out of sight.

The five skills close the gap in two movements. Orchestration, agentic collaboration and experimentation make agents productive. Trust and governance make them safe to scale. A leader who masters the first three without the last two simply scales risk faster.

Orchestration: designing the system of work

Orchestration is the skill of designing how work flows between people, agents and systems across the enterprise. It starts with a question most organizations have never asked explicitly: for each step of a value stream, should the work be done by a person, by an agent, or by both, and where do the hand-offs sit? Every hand-off between a human and an agent, or between two agents, is a trust boundary that must be designed.

Software development previews what is coming to all knowledge work. Agents increasingly write the code, while people specify, orchestrate and verify, and the constraint moves from production capacity to review capacity. Microsoft's 2025 Work Trend Index calls the resulting role the "agent boss" and warns about the human-agent ratio: past a certain point, more agents overwhelm the human capacity to apply judgment. Structures will change accordingly. Gartner predicted in 2024 that through 2026, 20% of organizations would use AI to flatten their structure , eliminating more than half of their current middle-management positions. In professional services, banking operations or customer service, the classic pyramid is becoming an orchestration model. The question is no longer how many people sit under each manager, but how much specification and review capacity each leader has.

Orchestration also means knowing which frictions in the business have just become solvable, and which problems are worth solving at all. It requires two kinds of fluency at once: agent orchestration, meaning systems design, model choice and verification at scale; and enterprise orchestration, meaning a sense of which problems carry value and which hand-offs create risk. Neither is a technical specialty. Both are now part of general management.

Agentic collaboration: working with agents every day

If orchestration designs the system, agentic collaboration is how a leader works inside it, day after day. It starts with specification. "Help our customers" is not a goal an agent can pursue. "Resolve billing disputes under R$500, within policy, in under ten minutes, and escalate anything else" is. Writing that sentence, with its outcome, constraints and escalation criteria, is the same discipline good leaders have always applied to delegating to people. It has simply become non-negotiable.

A useful image is onboarding a brilliant new hire. The prompt is what you tell them. The context is what is on their desk: the files, history and information they need. The harness is the office they work in: tools, permissions, workflows and review. Most organizations still polish the instructions while neglecting the desk and the office. The ladder extends further: once the harness exists, agents can be run in continuous loops and, eventually, improve themselves. At each level the leader's role changes, from author to curator, architect, operations designer and finally governor. Most organizations today sit at the first three levels.

Collaboration also requires calibrated trust. The jagged frontier study of 758 consultants found large gains on tasks inside the model's capability and markedly worse results just outside it, with no visible boundary between the two. A randomized trial by METR found experienced developers 19% slower with AI while believing they were 20% faster. Our intuitions are unreliable in both directions, so autonomy should be extended in steps: the agent informs, recommends, acts with approval, acts within a mandate, and only then anticipates.

Finally, agentic collaboration is never only with machines. The Afferolab finding that opens this article, that only 6.7% of leaders delegate while developing their teams, is a warning. The leader who uses agents to do everything personally has collaborated with the machine and abandoned the people. Excellence means delegating to agents and to humans at the same time, and using the time freed to develop both.

Experimentation: learning faster than the tools change

The tools change every few months. No leader can plan the agentic organization in advance; it has to be discovered through disciplined experiments. Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025 . Abandonment is not the problem. An experiment that ends quickly and cheaply is a success. The problem is scaling without evidence, and pilots that never end because nobody defined what success looked like.

Experimentation as a leadership skill has four disciplines. A clear hypothesis about which task or decision will change. A success metric agreed before the test begins. A comparison group, because without one, enthusiasm passes for evidence. And explicit kill criteria. Capital then follows the level of evidence reached: vendor demos justify curiosity; a retrospective test on the company's own data justifies a pilot; a live pilot against a comparison group justifies scaling in one unit; measured impact at scale justifies enterprise-wide investment.

Experimentation is also what the company does with the capacity AI frees. Gartner expects that by 2027, 75% of organizations that bank AI productivity gains as cost savings will be outpaced by competitors that reinvest them . Reinvesting freed time in new experiments, with customers, products and processes, is how the learning loop compounds. Experiments must also respect how reliability compounds. A process that is 95% reliable at each step succeeds end to end only about 36% of the time after 20 steps. Test agents on the company's own tasks, not on vendor demos, and break long flows into segments that can be verified one by one.

None of this works without psychological safety in Amy Edmondson's classic sense : people must be able to report a failed experiment without fear, or the organization will learn only from the experiments that flatter it.

AI tools aren't a creativity machine on their own. Jing Zhou , Rice University, 2025

Trust: the scarce currency in a world of infinite content

The same technology that multiplies what a leader can produce also multiplies what nobody has checked. Content, data and even decisions are now generated at a volume no organization can verify line by line. Left unmanaged, this abundance does not create shared understanding. It fragments it.

The warning signs are visible on three fronts. In the laboratory, Ilia Shumailov and colleagues showed in Nature that models trained indiscriminately on machine-generated content suffer irreversible defects, in which "tails of the original content distribution disappear." Organizations can decay in a similar way when unverified synthetic output feeds decisions that feed more output. The effect is already visible in public knowledge: volunteers for four African-language editions of Wikipedia estimated that 40% to 60% of their articles were uncorrected machine translations , and models trained on those pages get worse, producing more bad pages. The rule for any corporate knowledge base follows: nothing enters the corpus without an accountable reviewer. In the public sphere, the Reuters Institute's Digital News Report 2026 found trust in news at 37%, the lowest since 2015, and only 12% of people comfortable with news made entirely by AI . In business, deception has become industrial: the engineering group Arup lost US$25.6 million to a deepfake video call impersonating its CFO, and the FBI recorded nearly US$893 million in losses from AI-related complaints in 2025.

If you use machine translation responsibly, it can be efficient and useful. Unfortunately, you cannot trust all people to use it responsibly. Amir Aharoni , Wikimedia Language Committee, via MIT Technology Review , September 2025

People remain the preferred point of attack, and AI makes them easier to deceive. Verizon's 2026 Data Breach Investigations Report found the human element in 62% of breaches and third parties involved in nearly half, and observed that attackers now use AI mainly to accelerate familiar methods, from phishing to social engineering, rather than to invent new ones. When a familiar voice or face can be cloned, recognition is no longer proof of identity; verification has to travel through a second channel.

The barrier to attack has fallen with each era of technology: from scarce expertise, to tools that could be bought, to criminal services that could be rented, and now to prompts that anyone can type. The prompt era is documented. In November 2025 Anthropic disclosed that a group it assessed as Chinese state-sponsored had used an AI coding agent to run most of an espionage campaign against about thirty targets, with AI performing 80–90% of the work, and in 2026 Google's Threat Intelligence Group reported the first zero-day exploit it believes was developed with AI .

The deeper risk is social, and it is the part of this transition I insist on in everything I write. Uncontrolled, unverified and infinite data widens the chasms between people: between those who can tell the plausible from the true and those who cannot; between employees who master AI and those left behind; between the 2.2 billion people still offline and everyone else; and between groups fed different synthetic versions of reality. The International Monetary Fund estimates that about 60% of jobs in advanced economies are exposed to AI, against 26% in low-income countries .

Trust as a leadership skill works in two directions. Toward machines, it means calibrated reliance and preserved human judgment: Gartner expects that, through 2026, the atrophy of critical-thinking skills caused by generative AI will push half of global organizations to require "AI-free" skills assessments . Toward people, it means transparency about when agents are involved and how their outputs are used. Many employees already hide their use of AI. Ethan Mollick calls them "secret cyborgs" : hiding is individually rational when the person who discloses carries the risk and the firm captures the benefit. The remedy is a credible commitment that productivity gains will go to redeployment and growth, not only to cuts. Trust lost at machine speed is rebuilt only at human speed.

Machine-augmented humans who keep themselves hidden. Ethan Mollick , "Detecting the Secret Cyborgs", One Useful Thing , 2023

Governance: making autonomy safe to scale

Governance is often treated as the brake on AI. In the agentic era it is the opposite: it is what allows a leader to extend autonomy with confidence. Gartner defines AI TRiSM , AI trust, risk and security management, as the framework and technical capabilities that keep AI systems trustworthy, secure and compliant through continuous monitoring, validation and enforcement. It predicts that by 2030, guardian agents , AI that supervises other AI, will account for at least 10 to 15% of agentic AI markets, and that 80% of Global 500 companies will make their CIO or chief AI officer the contractual "evidence custodian" for AI accountability .

For the leader, governance translates into five practical rules. Every agent has a named human owner. Permissions are proportionate to consequences, and consequential actions keep a human approval point. Every agent leaves an audit trail of what it did and why. Content and data carry provenance, supported by standards such as C2PA content credentials and, in Europe, by the AI Act's transparency obligations for AI-generated content, whose grace period for existing systems ends in December 2026 under the digital omnibus . And accountability never transfers to the machine: when Air Canada argued that its chatbot was responsible for its own statements, the tribunal called it "a remarkable submission" .

Written rules are only half of governance. Shadow AI appears wherever policy and practice diverge. Most corporate AI policies are still written in the vocabulary of the duty ethic and the mechanical paradigm; they will be obeyed on paper and bypassed in practice. A leader governs well when the norms people actually live by, not only the policies they sign, match the level of autonomy the organization has chosen to grant.

The same lesson applies to agents themselves. Instructions in a prompt, such as "never delete data" or "always ask before sending", are soft guardrails: the model usually follows them, but it reads its orders and the content it processes as one stream of text, so hidden instructions in an email, a web page or a résumé can override them. Simon Willison calls the most dangerous configuration the lethal trifecta : an agent with access to private data, exposure to untrusted content and the ability to communicate externally. Only hard guardrails, enforced outside the model, reliably limit the damage: permissions, allow-lists, sandboxes, rate limits and human approval of the specific transaction, not just the session. The practical frame is to govern agents as we govern people with authority. Vet them before hiring, limit their access and separate their duties while they work, monitor for anomalies, and revoke their credentials when they leave. Soft guardrails lower the probability of harm; hard guardrails cap its impact.

Leaders also need a map of where an AI application can be attacked. The OWASP Top 10 for LLM Applications shows that the model is only one of the targets: an attacker can poison what it learns from, inject instructions through what it reads, steal it, flood it with costly requests, compromise the components it is assembled from, or exploit what happens after it answers, when output flows unchecked into another system or a plugin acts with more authority than anyone intended. The 2025 edition adds system-prompt leakage, weaknesses in the stores that feed retrieval, misinformation and unbounded consumption. The risk concentrates at the edge, where an agent turns a wrong answer into a wrong action, and that is exactly where hard guardrails belong.